A method and apparatus for AI-assisted knowledge assessment in adaptive learning

Abstract

This application provides an AI-assisted knowledge assessment method and apparatus for adaptive learning, effectively solving the problems that traditional classrooms and traditional intelligent tutoring systems cannot provide an adaptive learning environment and timely and effective student assessment. The method includes: pre-setting a learning task containing multiple knowledge skills in the digital courseware system; acquiring learning data from students' responses to the learning task; decomposing the learning task into multiple knowledge skills; mapping the multiple knowledge skills to the learning data; filtering out key learning data; inputting the key learning data into a student model to obtain corresponding assessment results; fusing multiple assessment results to obtain an assessment level result; using the assessment level result to assist in assessing the student's knowledge status; and generating a personalized learning path for the student's adaptive learning.

Description

Technical Field

[0001] This application relates to the field of educational assessment technology, and more specifically, to an artificial intelligence-assisted knowledge assessment method and apparatus for adaptive learning. Background Technology

[0002] Artificial intelligence (AI)-assisted knowledge assessment methods have moved from research labs into practical applications, providing adaptive learning environments in real-world classrooms (Baker, 2016; Romero & Ventura, 2020). Furthermore, some of these have been deployed in online education environments, providing high-quality education to students worldwide with time and cost savings (e.g., digital learning platforms, MOOCs, and other online education platforms). Learning alongside computers plays a vital role today and will become an inevitable trend in the future of education. However, providing adaptive learning experiences to every student in a cost-effective manner, and ensuring equal educational opportunities for every student worldwide, remains a challenge.

[0003] Adaptive learning is essential for improving individual learning outcomes and enhancing the student experience in personalized learning environments. It has been proven that adaptive learning environments are more effective than traditional classroom learning environments (Desmarais & d Baker, 2012). Adaptive learning environments need to address the challenges of large-scale personalization in the real-world human learning process. Various forms of tutoring systems are equipped with adaptive learning guidance, and some successful systems are used by tens of thousands of students annually, with the number of users continuing to grow.

[0004] Providing an adaptive learning environment for each student in traditional classrooms and traditional intelligent tutoring systems has always been a major challenge. Because each student has different knowledge bases, learning speeds, and learning styles, a uniform teaching model often fails to meet the needs of all students, leading to significantly reduced learning outcomes. Furthermore, extending adaptive learning models to large-scale teaching, especially for thousands or even millions of students, presents extremely high technical difficulties and computational costs. Traditional assessment methods in traditional classrooms and intelligent tutoring systems, whether teacher-led or system-based, often suffer from delays and errors. They cannot provide immediate feedback or recommend personalized learning materials, impacting students' learning experience and effectiveness. They also cannot adjust teaching strategies and content in real time based on students' knowledge levels or skills, resulting in rigid knowledge assessments and teaching outcomes that fail to meet students' individual learning needs. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide an AI-assisted knowledge assessment method and device for adaptive learning. This AI-assisted knowledge assessment method and device for adaptive learning effectively solves the problem that traditional classrooms and traditional intelligent tutoring systems cannot provide an adaptive learning environment and conduct timely and effective assessment of students.

[0006] In a first aspect, embodiments of this application provide an AI-assisted knowledge assessment method for adaptive learning, applied to a digital courseware system, the method comprising:

[0007] Learning tasks containing multiple knowledge and skills are pre-set in the digital courseware system, and learning data is obtained when students answer the learning tasks.

[0008] The learning task is decomposed to obtain multiple knowledge skills, the multiple knowledge skills are mapped to the learning data, and key learning data is filtered out from the learning data and input into the student model;

[0009] The key learning data is matched with multiple sub-models in the student model to obtain the corresponding evaluation results by performing knowledge status evaluation based on the matched key learning data.

[0010] Multiple assessment results are integrated to obtain assessment level results, which are used to assist in assessing students' knowledge status and generate personalized learning paths for students' adaptive learning.

[0011] In conjunction with the first aspect, this application provides a first possible implementation of the first aspect, wherein inputting the key learning data into the student model includes:

[0012] Based on different types of key learning data and mapped knowledge and skills, establish student response matrix, student profile matrix and skill mapping matrix respectively;

[0013] A student model is constructed based on the student response matrix, student profile matrix, and skill mapping matrix, and the key learning data is received after the evaluation is passed.

[0014] In conjunction with the first aspect, this application provides a second possible implementation of the first aspect, wherein the key learning data includes dynamic data;

[0015] The process of establishing student response matrices, student profile matrices, and skill mapping matrices based on different types of key learning data and mapped knowledge and skills includes:

[0016] Extract the timestamps from the dynamic data, and process the dynamic data based on the timestamps to obtain dynamic sequence data;

[0017] Based on the dynamic sequence data, student response matrix, student profile matrix, and skill mapping matrix are established respectively.

[0018] In conjunction with the first aspect, this application provides a third possible implementation of the first aspect, wherein the step of filtering key learning data from the learning data includes:

[0019] The correlation between the learning data and the knowledge state is calculated using a pre-set correlation calculation network, and it is determined whether the correlation meets the preset correlation conditions.

[0020] If so, then the learning data is determined to be key learning data.

[0021] In conjunction with the first aspect, this application provides a fourth possible implementation of the first aspect, wherein the fusion of multiple evaluation results to obtain the evaluation level result includes:

[0022] Based on the type of the learning task and the relationship between the evaluation results, the fusion dimension of the evaluation results is determined;

[0023] The preset fusion method of the fusion dimension is invoked to fuse the multiple evaluation results to obtain the evaluation level result.

[0024] In conjunction with the first aspect, this application provides a fifth possible implementation of the first aspect, wherein the fusion dimension includes a collaborative dimension;

[0025] The step of invoking the preset fusion method of the fusion dimension to fuse the multiple evaluation results to obtain the evaluation level result includes:

[0026] Based on the knowledge and skills examined in the learning task, determine whether the relationship between the knowledge and skills meets the conditions for synergistic integration;

[0027] If so, the multiple evaluation results are logically combined to obtain a comprehensive judgment result, and an evaluation level result is obtained based on the comprehensive judgment result.

[0028] In conjunction with the first aspect, this application provides a sixth possible implementation of the first aspect, wherein matching the key learning data with the sub-models in the student model includes:

[0029] According to the predefined attribute partitioning relationship, the key learning data is partitioned into the corresponding attribute types;

[0030] Based on the sub-model type of the student model matched by the attribute type, determine the sub-model matched by the key learning data.

[0031] Secondly, embodiments of this application provide an AI-assisted knowledge assessment device for adaptive learning, applied to a digital courseware system, the device comprising:

[0032] The setting module is used to pre-set learning tasks containing multiple knowledge and skills in the digital courseware system and obtain learning data when students answer the learning tasks.

[0033] The decomposition module is used to decompose the learning task into multiple knowledge skills, map the multiple knowledge skills to the learning data, and filter out key learning data from the learning data to input the key learning data into the student model.

[0034] The evaluation module is used to match the key learning data with multiple sub-models in the student model, so as to perform knowledge status evaluation based on the matched key learning data and obtain the corresponding evaluation results.

[0035] The generation module is used to integrate multiple assessment results to obtain assessment level results, and to assist in assessing students' knowledge status based on the assessment level results, and to generate personalized learning paths for students' adaptive learning.

[0036] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of any one of the artificial intelligence-based assisted knowledge assessment methods described above.

[0037] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of any one of the artificial intelligence-based assisted knowledge assessment methods described above.

[0038] This application provides an AI-assisted knowledge assessment method for adaptive learning, applied to a digital courseware system. The method first pre-sets a learning task containing multiple knowledge skills within the digital courseware system and acquires learning data from students' responses to the learning task. Next, it decomposes the learning task into multiple knowledge skills, maps these skills to the learning data, and filters out key learning data from the learning data to input into a student model. Then, it matches the key learning data with multiple sub-models within the student model to perform knowledge status assessment based on the matched key learning data, obtaining corresponding assessment results. Finally, it integrates multiple... The assessment results yield evaluation level results, which are used to assist in assessing students' knowledge status and generate personalized learning paths for students. This provides students with an adaptive learning environment within the digital courseware, accurately assesses their knowledge status, and provides personalized learning paths based on this assessment. This achieves efficient adaptive learning and effectively solves the problems of traditional classrooms and traditional intelligent tutoring systems failing to provide an adaptive learning environment and timely and effective assessment of students. It also ensures accurate assessment of students' knowledge status, the accuracy and effectiveness of the personalized learning paths generated for students, and the auxiliary role in students' learning. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart illustrating the first AI-assisted knowledge evaluation method for adaptive learning provided in an embodiment of this application is shown.

[0041] Figure 2 A schematic diagram of static data provided in an embodiment of this application is shown;

[0042] Figure 3 A schematic diagram of the student model provided in an embodiment of this application is shown;

[0043] Figure 4 This illustration shows a schematic diagram of dynamic data provided in an embodiment of this application;

[0044] Figure 5 A schematic diagram of the hidden Markov model provided in an embodiment of this application is shown;

[0045] Figure 6 This paper illustrates a structural block diagram of a first AI-assisted knowledge evaluation method for adaptive learning provided in an embodiment of this application.

[0046] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0048] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0049] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0050] Providing an adaptive learning environment for each student remains a significant challenge in traditional classrooms and traditional intelligent tutoring systems. Because each student has different knowledge bases, learning speeds, and learning styles, a uniform teaching model often fails to meet the needs of all students, leading to a significant reduction in learning effectiveness. Furthermore, extending adaptive learning models to large-scale teaching, especially for thousands or even millions of students, presents extremely high technical difficulties and computational costs. Traditional assessment methods in traditional classrooms and intelligent tutoring systems, whether teacher-led or system-based, often suffer from delays and errors. They cannot provide immediate feedback or recommend personalized learning materials, impacting students' learning experience and outcomes. They also cannot adjust teaching strategies and content in real time based on students' knowledge levels or skills, resulting in rigid knowledge assessments and teaching effectiveness, failing to meet students' personalized learning needs.

[0051] Based on this, this application provides an AI-assisted knowledge evaluation method and apparatus for adaptive learning, which will be described below through embodiments.

[0052] Example 1

[0053] To facilitate understanding of this embodiment, a detailed description of an AI-assisted knowledge evaluation method for adaptive learning disclosed in this application embodiment will be provided first. For example... Figure 1 The flowchart shown illustrates an AI-assisted knowledge evaluation method for adaptive learning. This application provides an AI-assisted knowledge evaluation method for adaptive learning, applied to a digital courseware system. The method includes:

[0054] S101. Pre-set learning tasks containing multiple knowledge and skills in the digital courseware system, and obtain learning data when students answer the learning tasks;

[0055] S102. Decompose the learning task to obtain multiple knowledge skills, map the multiple knowledge skills to the learning data, and filter out key learning data from the learning data to input the key learning data into the student model;

[0056] S103. Match the key learning data with multiple sub-models in the student model to obtain the corresponding evaluation results by performing knowledge status evaluation based on the matched key learning data.

[0057] S104. Integrate multiple assessment results to obtain assessment level results, and use the assessment level results to assist in assessing the student's knowledge status and generate a personalized learning path for the student's adaptive learning.

[0058] In step S101, when students use the digital courseware system for knowledge learning, the system can automatically generate or set learning tasks for students based on pre-filled information such as age, grade, or knowledge learned. These learning tasks are adaptive learning tasks that include multiple knowledge skills, generated according to the student's specific situation. For example, if the pre-set learning task is "1+2*3.5=?", the knowledge skills tested are integer addition, decimal multiplication, decimal representation, decimal multiplication, decimal addition, etc. That is, knowledge skills are a fine-grained representation of the learning task. A learning task can be decomposed into multiple knowledge skills or into one knowledge skill, depending on the learning tasks set by the digital courseware system, or it can be set based on the learning requirements of teachers, parents, and the student themselves.

[0059] The digital courseware system records learning data generated by students when answering learning tasks. This learning data includes various elements, such as the number of times a student answers a question, the answering time, and the model of the terminal on which the digital courseware system is located. All of this learning data is stored in the digital courseware system. To ensure efficient processing of the learning data, it is preprocessed. This preprocessing mainly includes data cleaning, format conversion, and outlier handling. For example, duplicate or invalid answer records, such as those with excessively short dwell times or abnormal answer patterns, are removed; unstructured data, such as free text feedback, is converted into a structured format; and missing values ​​in the learning data are imputed or marked.

[0060] The digital and intelligent courseware system is an advanced teaching resource that integrates digital and intelligent technologies. Built on digital technology, it transforms traditional textbooks, supplementary materials, and various educational resources into a computer-processable digital format, facilitating storage, management, and dissemination. Digital and intelligent courseware utilizes multimedia, animation, and simulation to provide a highly interactive learning experience, enabling students to understand and master knowledge through dynamic visualization, enhancing participation and engagement in the learning process. Furthermore, combining artificial intelligence and big data analytics, it can dynamically adjust and personalize content based on learners' learning behaviors, ability levels, and individual needs, achieving adaptive teaching. It also features real-time learning data collection and analysis capabilities, providing feedback and evaluation to help curriculum designers optimize teaching strategies and guide students in improving their learning methods. In addition, it supports students' self-directed and exploratory learning, cultivating their innovative thinking and independent problem-solving abilities. Digital and intelligent courseware aims to empower education through technology, improve teaching quality, promote efficient learning, and meet the development trends of modern educational informatization and personalized teaching.

[0061] In step S102, the learning task is decomposed to obtain multiple knowledge skills examined by the learning task. Each knowledge skill is a knowledge component KCs, that is, the knowledge skills such as integer addition, decimal multiplication, decimal representation, decimal multiplication, and decimal addition obtained in step S101 are all knowledge components KCs. After obtaining the multiple knowledge skills, they are mapped to the learning data according to a preset mapping method, and key learning data filtered from the learning data is input into the student model. The key learning data refers to learning data that can affect the assessment of the student's knowledge status, while learning data that is irrelevant to the assessment of the student's knowledge status is filtered out, thereby ensuring the accuracy of the assessment of the student's knowledge status and reducing the amount of data processing. The student model has multiple sub-models; such as Figure 3 As shown, the student model is composed of multiple sub-models. The key learning data is output to multiple sub-models. The sub-models include a diagnostic classification model, an item response theory model, a hidden Markov model, a factor analysis model, and a deep learning model. The diagnostic classification model includes compensatory and non-compensatory models. The student model is trained on historical key learning data and has been evaluated and passed, thereby ensuring the accuracy and effectiveness in processing the key learning data and ensuring the accuracy of the knowledge state obtained through assistance.

[0062] In a specific implementation of step S102, one embodiment is as follows: the step of filtering key learning data from the learning data includes:

[0063] S10211. Calculate the correlation between the learning data and the knowledge state using a preset correlation calculation network, and determine whether the correlation meets the preset correlation conditions.

[0064] S10212. If so, then the learning data is determined to be key learning data.

[0065] In steps S10211-S10212, to avoid excessive learning data that might prevent the student model from quickly receiving all the learning data and thus affect the knowledge state assessment based on the learning data, feature filtering is performed on the learning data. Specifically, a pre-set correlation calculation network is used to calculate the correlation between the learning data and the knowledge state, and it is determined whether the correlation meets a preset correlation condition. The pre-set correlation calculation network can be a correlation analysis algorithm or an information gain algorithm. The preset correlation condition is that the correlation is higher than a preset correlation threshold, which is determined based on the actual situation. If the preset correlation condition is not met, the learning data is deleted, such as the terminal type of the digital courseware system. If the preset correlation condition is met, the learning data is used as key learning data, such as the error rate of specific knowledge skills, the response time of learning tasks, etc., and the key learning data is input into the student model after the assessment is completed. The core of filtering the key learning data is to mine potential factors related to knowledge components (KCs) or skills from the learning data. For example, by extracting students' knowledge forgetting curves through time series analysis, or by mining the dependencies between knowledge points through association rules (such as the impact of mastering "fraction operations" on "algebraic equations"), these features will be input into learning models (such as RNNs and Bayesian networks) to assess students' knowledge status and predict their learning trajectories.

[0066] In a specific implementation of step S102, another embodiment exists: inputting the key learning data into the student model includes:

[0067] S1021. Based on different types of key learning data and mapped knowledge and skills, establish student response matrix, student profile matrix and skill mapping matrix respectively;

[0068] S1022. Construct a student model for the student based on the student response matrix, student profile matrix and skill mapping matrix, and receive the key learning data after the evaluation is passed.

[0069] In steps S1021-S1022, for ease of processing, the key learning data is divided into static data and dynamic data according to its type. Static data refers to the test results or evaluation results obtained by students when completing learning tasks at a specific point in time. This type of static data only reflects the student's current knowledge state. For example, a student's answer result (correct or incorrect) in a unit test is recorded as a binary value (1 indicates mastery, 0 indicates no mastery). However, "learning 1" and "no learning 0" here do not directly correspond to the definition of static data, but rather refer to the binary representation of the knowledge state. Dynamic data, on the other hand, refers to the real-time behavioral data generated by the student when answering this learning task, such as the student's answer sequence, interaction duration, number of error corrections, and knowledge point jump paths. This learning data has temporal sequence and continuity, and can reflect the dynamic changes in the student's knowledge state, such as the process of trying the same question multiple times or the transfer behavior between different knowledge points.

[0070] In the static data dimension, assuming "no learning" occurs during the test, and therefore the students' knowledge status remains unchanged, this static data is also called cross-sectional data. It appears in the form of a binary matrix called the student response matrix R. The student response matrix R represents the students' outcome data, where success is assigned 1 and failure is assigned 0. The following is a student response matrix R composed of the outcome data of 4 students and 5 learning tasks:

[0071]

[0072] Modeling with this type of data is typically done in the context of summative assessment, which aims to evaluate student learning at the end of a teaching unit by comparing them to certain standards or benchmarks. Information from summative assessments can then be used to guide their activities in subsequent lessons. This type of data often has only a few or no missing responses from respondents (students), and in responding to the learning task, each student must answer the same number of questions in the same order, such as... Figure 2 As shown, the first row contains the learning task number, and the second row contains the result data.

[0073] The skill mapping matrix Q represents the mapping from learning tasks to knowledge skills. The following is the skill mapping matrix Q for 5 learning tasks and 3 knowledge skills:

[0074]

[0075] The skill mapping matrix Q not only defines the direct association between each learning task and its knowledge components KCs, but also represents complex tasks requiring the collaborative application of multiple skills by introducing implicit integration nodes or higher-order skill dimensions. For example, for the problem "1 + 2 × 3.5", the Q matrix not only marks the required basic skills such as "integer addition" and "decimal multiplication", but also associates an independent integrative skill such as "arithmetic expression priority handling". This application adds a column for integrative skills to the traditional Q matrix. That is, whether an integrative skill is triggered depends on whether the combination of basic skills satisfies a specific logic, such as "all basic skills must be mastered" or "at least one skill must be mastered". For example, if a problem requires skills A and B, then its corresponding integrative skill C is marked as 1 in the Q matrix, and the mastery status of C is determined by the joint state of A and B.

[0076] The student file matrix A is then shown below:

[0077]

[0078] The probability of answering a question correctly depends primarily on the knowledge state of the knowledge skills behind the learning task, where knowledge state refers to knowledge level. Furthermore, it may require a specific skill to correctly integrate all skills. In the Knowledge-Learning-Guided (KLI) framework (Koedinger & Anderson, 1998), this specific skill is defined as an "integrative knowledge component" that integrates with all other KCs to produce the correct answer. Therefore, a student model is established based on different types of learning data, including student response moments R, student profile matrix A, and skill mapping moments Q. This student model is represented by the following Boolean matrix product:

[0079] R = A⊙R (4).

[0080] Based on the student model above, key learning data can be received to assess the knowledge status of students. In some models, the values ​​of A and R may both be real values ​​in the interval [0,1]. If a standard dot product is used, such real values ​​may appear in D, and the product of R = A ⊙ Q may even produce values ​​outside the range [0,1]. In this case, the result is usually rounded to the binary value {0,1}.

[0081] In a specific implementation of step S10221, one embodiment is as follows: the key learning data includes dynamic data;

[0082] The process of establishing student response matrices, student profile matrices, and skill mapping matrices based on different types of key learning data and mapped knowledge and skills includes:

[0083] S102211. Extract the timestamp of the dynamic data, and process the dynamic data based on the timestamp to obtain dynamic sequence data;

[0084] S102212. Based on the dynamic sequence data, establish a student response matrix, a student profile matrix, and a skill mapping matrix respectively.

[0085] In steps S102211-S102212, within the dynamic data dimension, dynamic data not only records students' specific behaviors during the learning process (such as answer results and interactive actions), but also precisely records the temporal information of these behaviors. For example, the system marks the start time, submission time, intervals between multiple attempts on the same question, and the time trajectory of jumping between different knowledge points for each student's answer. This embedding of timestamps allows dynamic data to be presented in a sequential form (e.g., "Question A is answered incorrectly at time t1 → Question B is answered correctly at time t2 → Question A is retried at time t3"), thus fully depicting the dynamic evolution of students' knowledge state. The assumption of "no learning" in steps S1021-S1022 is unrealistic for the digital courseware system because students learn while using it. In this case, students may attempt the same type of question multiple times, potentially for a long time. The system records all student behaviors throughout the entire learning phase. Dynamic data necessarily involves the concept of sequence, and is usually... Figure 4 The winning bid has a timestamp, such as Figure 4 As shown, the first row represents the number of interactions students made in responding to learning tasks, and the second row represents the result data. For example, the number "2" represents that the student performed two actions on knowledge point A (such as answering questions, viewing explanations, and repeating exercises), while "5" represents five interactions on knowledge point B. This type of data is also known as longitudinal data. Modeling with this type of data is for performing knowledge tracing and is typically used in the context of formative assessment. Depending on the nature of the tutoring system, this data often has many missing values. Therefore, the dynamic data needs to be imputed or labeled, and the dynamic data also needs to be normalized. Normalization standardizes data of different dimensions or ranges to a uniform interval (such as [0,1]). For example, the time spent answering questions can be converted into a relative time percentage, or the student's mastery of a certain skill can be quantified into a continuous value between 0 and 1 through a probability model (instead of simple binarization). However, in certain models (such as DINA / DINO), the skill mastery status may be binarized (1 / 0), but this is a model parameter setting, not a general operation of normalization. The student response moment R, student profile matrix A, and skill mapping moment Q based on dynamic sequence data are the same as in steps S1021-S1022.

[0086] This application refers to the model combining the student profile matrix A and the skill mapping matrix Q as a diagnostic classification model. Many models originate from psychometrics. Compared to the classic Item Response Theory (IRT), which covers continuous latent attributes, the Diagnostic Classification Model (DCM) represents discrete latent attributes. These attribute patterns are binary vectors, where 1 indicates mastery of the latent attribute and 0 indicates the opposite. These patterns provide feedback to teachers to help design remedial instruction. They also have other names: restricted latent category model, latent response model, multiclass latent category model, cognitive diagnostic model, cognitive psychometric model, structured item response model, and structured localized latent category model. Various DCMs have been proposed in the past decade. Based on the nature of the model, DCMs can be divided into two main categories: compensatory and non-compensatory. Non-compensatory DCM models assume that each item requires a specific set of knowledge and skills, and the lack of any one of them may lead to student failure. This type of model is called the Deterministic Input Noise and Effect (DINA) model. It is a well-known non-compensatory model. In contrast, the Noise-Or-Gate (DINO) model represents the other extreme of this requirement: any skill is sufficient for success. The DINA model is also considered a joint model, while DINO is considered a disjunctive model. Regardless of their formulations, all the models mentioned above share a fundamental idea: associating each learning task with one or more knowledge skills.

[0087] Compensatory models posit that any attribute can trigger an affirmative response to a claim, but not necessarily. However, the simultaneous presence of two attributes is more likely to trigger an affirmative response than either attribute alone. A similar argument can be made for two knowledge skills associated with a learning task. Mastering either of the two knowledge skills may lead to success on that learning task, but mastering both knowledge skills will make success more likely. The Noise-Or-Gate (DINO) model is a well-known disjunctive model in skill modeling. The DINO model is a fully compensatory DCM. Students have a high probability of providing the correct answer as long as they possess any one of the required knowledge skills, rather than requiring all of them. Given a response matrix, this application will evaluate the knowledge skills or attributes underlying those learning tasks. An evaluation consisting of I learning tasks considers a domain that measures K attributes or knowledge skills.

[0088] Let X ij i = 1, 2, ..., I j = 1, 2, ..., J represents the binary 0 / 1 response of student i to learning task j, where 1 represents that the student provided a correct response to the learning task, and 0 represents otherwise. The knowledge and skills α of student i are also represented. iIt is a binary vector of length K, consisting of 0 and 1 elements, where 1 indicates that the student has mastered the knowledge or skill, and 0 indicates otherwise. For a learning task requiring K knowledge or skills, students can be classified into 2 categories. K One of the possible knowledge and skills. P ij Let be the probability that student i correctly answers learning task j, given by the DINO model as follows:

[0089]

[0090] in:

[0091]

[0092] s j =P(X) ij =0|ξ ij =1)

[0093] g j =P(X) ij =1|ξ ij =0);

[0094] Where K represents knowledge and skills, and X... ij It is student i's response (result) to learning task j, q jk It is the (j, k)th node of the skill mapping matrix Q. th elements, α ik This is the attribute pattern of student i. The model is composed of s j Slip (the probability that a student who has mastered all the required knowledge and skills will give an incorrect answer to learning task j) and g j The probability of guessing (the probability that a student who has not mastered any of the required skills will answer learning task j correctly) is parameterized. According to the formula above, the probability of answering learning task j correctly has two possibilities. If the student has not mastered any of the required attributes, it is still possible to provide a correct answer through guessing; therefore, in this case, the probability of a correct response is g. j When a student has mastered at least one of the required knowledge and skills, the probability of a correct response is 1 - s. j The DINO model can be estimated using Markov chain Monte Carlo (MCMC) or as a constrained log-linear model with latent classes.

[0095] The most popular uncompensated DCM model is DINA, or Noise AND Gate model. Similar to the DINO model, the DINA model also considers the possibility that a respondent with all the required skills may miss an item, the possibility of careless error (glide parameter), and the possibility that a respondent lacking at least one required skill may give the correct response by guessing (guess parameter).

[0096]

[0097] in

[0098]

[0099] s j =P(X) ij =0|ξ ij =1)

[0100] g j =P(X) ij =1|ξ ij =0);

[0101] Where K represents knowledge and skills, and X... ij It is student i's response (result) to learning task j, q jk It is the (j, k)th node of the skill mapping matrix Q. th elements, α ik This is the attribute pattern of student i. The model is composed of s j Slip (the probability that a student who has mastered all the required knowledge and skills will give an incorrect answer to learning task j) and g j The conjecture (the probability that a student without any of the required knowledge and skills will correctly answer learning task j) is parameterized. The response outcome is binary: {0, 1}, as ξ... ij When = 1, student i has mastered all the required skills; otherwise, it is 0.

[0102] For decades, Item Response Theory (IRT) has been the foundation of computerized adaptive testing environment evaluation mechanisms. In IRT, the probability of successfully completing a learning task increases with proficiency (knowledge ability), assuming that a student's knowledge state is static during testing. When testing, the student's knowledge state θ is assessed based on their proficiency, and each test's learning task contributes information to refine the estimation of knowledge state. Beyond the static knowledge assumption, the original IRT model assumes a single skill and considers the test learning task to be one-dimensional.

[0103] The Rasch model is an example of IRT, demonstrating a strong theoretical foundation in both psychometrics and mathematical frameworks. IRT accepts binary item response outcomes and assigns a proficiency level θ to student i. i Measurements can be taken after each question. Each learning task j has its own difficulty β. j The main idea of ​​IRT is to use students' abilities and the difficulty of the learning task to estimate the probability that student i will correctly answer item j:

[0104]

[0105] Where σ is a logical function, θ i β represents student i's proficiency on the test learning task. j This represents the difficulty of learning task j. Multidimensional IRT models (a variant of IRT) can handle two or more dimensions, but their complexity is much greater, and their application in personalized learning environments is not yet widespread. An important consideration is that IRT is not considered a knowledge-tracking method because it assumes that students do not learn during testing. It is considered a knowledge assessment method, where each new test's learning task helps bring in information to refine student i's knowledge state θ. i The estimation. Note that when IRT is used in a learning environment where the expected state of knowledge will change, the evaluation is typically performed on the first attempt at a project, within a practice context that allows for multiple attempts. Alternatively, multiple attempts can also be calculated as different projects, and the β value corresponding to the attempt is used. j Use an index t.

[0106] Wilson et al. proposed a Bayesian extension of IRT for use in the context of dynamic data (this invention refers to this model as IRT*). IRT* uses a Bayesian approach and, by applying a formula to each θ... i and β j Apply independent standard normal prior distributions to logP ij Perform regularization. It maximizes the {θ} of the given response data {r:(i,j,r,t)∈D}. i ,β j The log-posterior probability of}, where the response r∈{0,1}, t is the time of each attempt, and the student response data D is a set of tuples (i,j,r,t) indicating the student, learning task, response, and response time for each response.

[0107]

[0108] IRT* utilizes information from directly interacting projects and students within the system. θ is calculated using the Newton-Raphson method. i and β j The method uses maximum a posteriori (MAP) estimation. This model is termed IRT*. IRT* demonstrates performance comparable to deep knowledge tracing in predicting student performance.

[0109] Bayesian Knowledge Tracing (BKT) was the earliest method used to simulate learners' changing knowledge states, arguably the first model to relax the assumption of static knowledge states. In the original BKT, each item tests a single skill, and the learner's skill mastery state is inferred at each time step based on a sequence of previous results. This method is particularly relevant for coaching systems that use practice and scaffolding as primary learning tools and monitor fine-grained skill mastery to determine the next step. It relies on Markov models to infer mastery states, from "not learned" to "learned," and... Figure 5 The above probability depends on the fixed parameters and the state at time step t.

[0110] BKT has four parameters, among which The prior probability that a student will master skill k; P(L) K P(G) represents the probability that a student who currently does not possess skill k will acquire it after the next practice opportunity; K P(S) represents the probability that a student guesses a learning task and gets the correct answer, even though they haven't mastered the skill k(guess); K ) represents the probability (slip) of a student answering a question incorrectly despite having mastered skill k.

[0111] In a typical learning environment with BKT, the student's assessment of their skill mastery is continuously updated each time they respond to a project. It is the probability that a student will master skill K at time t+1. Skill mastery is calculated as follows:

[0112]

[0113] P ij The probability that student i correctly applies skill k to solve project j at time t+1 is predicted based on the student's skill mastery at time t+1 as follows:

[0114]

[0115] First, the probability that a student has mastered knowledge skill K at timet t+1 is updated based on the student's observed Obs using Equation 7 in combination with Equation 5 or 6, where Obs ∈ {correct, incorrect}, representing the response evidence at timet t. Then, Equation 8 is used to estimate the likelihood that the student correctly applies knowledge skill K to solve learning task j at time t+1. When using the Bayesian knowledge tracing algorithm, if the system does not identify that the student has sufficient knowledge skill for that skill (e.g., the probability that the student has mastered the knowledge skill is less than 95%), the student is instructed to practice questions related to that specific skill.

[0116] Piech et al. introduced Deep Knowledge Tracking (DKT) in 2015. Similar to BKT, it uses data on trial skills and leverages performance results to predict future sequence attempts. DKT uses a large number of artificial neurons to represent latent knowledge states and incorporates a temporal dynamic structure, allowing the model to learn a student's knowledge state from the data. It encodes knowledge skills and student response attempts into a one-bit hot feature input vector, which is used as input at each timestamp t. The output layer y... t It provides the predicted probability, that is, the student will correctly answer that specific learning task at time t+1. Recurrent Neural Networks (RNNs) take the input vector sequence x1, ..., x... T Mapped to the output vector sequence y1, ..., y T Based on the hidden state sequence h1, ..., h T This is a continuous encoding of relevant information observed in the past. It is defined by the following equation:

[0117] h t =tanh(W hx x t +W hh h t-1 +b h ) Formula 9;

[0118] y t =σ(W yh h t -b y ) Formula 10;

[0119] h t =tanh(W hx x t +W hh h t-1 +b h ) Formula 11;

[0120] y t =σ(W yh h t +b y ) Formula 12;

[0121] In DKT, both the tanh and sigmoid functions are applied element-wise, and are passed through the input weight matrix W. hx Recursive weight matrix W hh Initial state h0, hidden state ht, and output weight matrix W yh Parameterization is performed. The biases of the latent and output cells are respectively determined by b. h and b y Indicates. x t It is student interaction xt ={k t r t A one-hot encoded vector representing the knowledge and skill k practiced. t Student response r t The combination of x, so x t ∈{0,1} 2M Based on the number M of unique skills, the output y t It is a vector of knowledge and skills number K, where each value represents the probability that a student correctly answers a question related to knowledge and skills K at time t+1.

[0122] Therefore, the probability that student i correctly answers the learning task j related to knowledge and skill k at time t+1 can be derived from the vector y. t The following results were retrieved from the search results:

[0123] P ij =P ij (k t+1 )∈y t Formula 13.

[0124] DKT uses recurrent neural networks (RNNs) to represent a student's latent knowledge space and the dynamic number of practice sessions. By leveraging a student's past performance history, it is possible to infer the knowledge growth a student has gained through a task.

[0125] This application also sets up cross-validation, grid search, and optimization effects and practical applications for the student model to ensure its effectiveness. Cross-validation, in particular, enables data-driven verification of generalization capabilities.

[0126] Students' historical learning data (such as answer records and interactive behaviors) are divided into multiple mutually exclusive subsets (e.g., 5-fold cross-validation) based on time or knowledge points. For example, data from a semester can be divided into a training set (4 subsets) and a validation set (1 subset), with the validation set being rotated to cover all data. After each training iteration, model metrics (such as AUC and RMSE) are calculated using the validation set to evaluate its generalization ability. For example, if the LSTM model experiences a sudden increase in prediction error rate during a certain fold of validation, it may indicate a special pattern in the data during that period (such as abnormal behavior caused by exam stress), requiring targeted optimization. Based on the cross-validation results, weaknesses in the model are identified (such as inaccurate predictions of "integrated skills"), and relevant training samples are added or feature weights are adjusted accordingly.

[0127] Grid search enables systematic optimization of the parameter space: First, it sets the search range for key hyperparameters for different sub-models. For example: number of hidden layers (1-3 layers), time step (10-50 steps), dropout rate (0.2-0.5); prior distribution type (Gaussian / Bernoulli), smoothing factor (0.1-1.0); glide parameter (0.01-0.2), guess parameter (0.05-0.3); it automatically traverses all parameter combinations, trains the model, and records performance. For example, it trains an LSTM with 3 hidden layers + 30 steps + 0.3 dropout rate, comparing its prediction performance with 2 layers + 50 steps + 0.2 dropout rate; the final parameters are selected based on validation set metrics (such as highest overall AUC, lowest RMSE). For example, if a set of parameters performs stably in multi-fold cross-validation, it is considered globally optimal.

[0128] The optimization and practical application ensure that the student model can be successfully applied in real-world scenarios. Cross-validation eliminates the risk of overfitting, ensuring the model maintains high accuracy even with new student data or across different knowledge points. For example, the optimized RNN model reduced the error rate from 15% to 8% when predicting the probability of mastering the "algebraic equations" skill. Grid search makes the model parameters more closely match real learning patterns (such as forgetting curves and skill transfer dependencies). For example, the adjusted DINO model more accurately reflects the compensatory logic that "mastering any skill is sufficient to solve problems," reducing misjudgments of complex learning tasks. The system periodically (e.g., monthly) re-executes cross-validation and grid search, dynamically updating the model based on new data to adapt to changes in teaching strategies or student groups.

[0129] After the sub-models in the student model are trained, the student model itself needs to be evaluated. Evaluating the sub-models is necessary to ensure that the student model can accurately assess the student's knowledge state during interaction with the learning system. In summative assessment, the knowledge and skills a student learns by solving a given learning task are just as important as predicting any other problems related to that skill. For example, IRT measures this using statistics such as model fit (Khalid). In formative assessment, the student model is typically validated by referencing two criteria. The first method of evaluating the student model is to measure the accuracy of predicting future student performance within the learning system. The second method is to validate using external metrics (e.g., the student's knowledge gain in a post-test). In both assessment scenarios, the student's actual knowledge state is latent and unobservable, but can be estimated based on their past performance on other items. The estimated knowledge state for a skill is used to predict the student's performance (success or failure) on the next problem related to that skill. Therefore, student models are often compared to their predictability in student performance. Thus, student models are typically evaluated based on the area under the receiver operating characteristic curve (AUC) and root mean square error (RMSE). AUC measures the entire two-dimensional area under the ROC curve, from (0,0) to (1,1), used for prediction, where the estimated value is the probability of the question being answered correctly. An AUC of 0.50 represents a score achievable by random guessing. Higher AUC scores indicate higher accuracy. RMSE is used as a variant of the Brier score to evaluate model fit. Lower RMSE values ​​indicate better model performance. AUC and RMSE provide robust metrics for evaluation and are commonly measured in knowledge-tracking and student performance prediction tasks.

[0130] In step S103, this application sets up multiple sub-models in the student model, and different sub-models have different effects when evaluating knowledge status for different key learning data. Therefore, after the student model receives the key learning data, it matches the key learning data with multiple sub-models in the student model so that the sub-models can evaluate knowledge status based on the matched key learning data to obtain evaluation results, thereby ensuring the accuracy of the obtained evaluation results and maximizing the reflection of the student's knowledge status of knowledge and skills. That is, different sub-models can obtain different evaluation results when evaluating knowledge status based on the matched key learning data, so there are multiple evaluation results. That is, this application designs a multi-model collaborative prediction mechanism in which the key learning data will be input in parallel and distributed according to the strengths of different sub-models, rather than being dominated by a single model.

[0131] In a specific implementation of step S103, one embodiment is as follows: matching the key learning data with the sub-models in the student model includes:

[0132] S1031. According to the predefined attribute partitioning relationship, partition the key learning data into the corresponding attribute types;

[0133] S1032. Based on the sub-model type of the student model matched by the attribute type, determine the sub-model matched by the key learning data.

[0134] In steps S1031-S1032, the predefined attribute division relationship refers to the static data and dynamic data included in the learning data. According to this attribute division relationship, the multiple data included in the key learning data are also divided into static data and dynamic data. This application matches the corresponding sub-model type of the student model for different attribute types. That is, static data corresponds to models such as logistic regression and Bayesian network for rapid diagnosis and probability inference; while dynamic data is preferentially input into deep learning models such as RNN and LSTM to capture the continuous evolution of knowledge state. If a learning task includes multiple knowledge skills, a cognitive diagnostic model based on the skill mapping matrix Q (such as DINA / DINO) is additionally called to ensure the accurate mapping driven by rules, thereby realizing targeted knowledge state evaluation of the key learning data to obtain the corresponding evaluation results.

[0135] In step S104, after obtaining the corresponding evaluation results from the predicted knowledge states of multiple sub-models, the student model needs to be given a specific evaluation level result. Therefore, it is necessary to fuse multiple evaluation results in a predetermined manner to obtain an evaluation level result. This evaluation level result is used to assist in evaluating the student's knowledge status regarding multiple knowledge skills in the learning task. The system then controls the digital courseware system to generate personalized learning paths for students based on the evaluation level result, thereby improving learning effectiveness and enhancing students' learning motivation and enthusiasm. The learning content recommendation module in the digital courseware system utilizes collaborative filtering and matrix factorization techniques, combined with the student's learning history and knowledge status, to provide personalized learning recommendations. The system integrates students' answers and learning behaviors in real time, dynamically adjusting recommendation strategies to ensure students receive learning resources that better match their learning needs and knowledge levels, thereby improving learning efficiency and effectiveness. The intelligent courseware system analyzes and provides feedback on students' learning behaviors and knowledge status, dynamically adjusting the student model to achieve highly adaptive learning capabilities. The system's instant feedback and interaction modules capture students' learning operations and answers in real time, feeding this information back to the system for dynamic adjustments. The system's self-improvement module continuously collects and analyzes real-time learning data to update and optimize the student model, achieving the most accurate assessment results. The digital courseware system of this application also utilizes big data analysis and data mining technologies to process and analyze learning data. By optimizing the architecture of the digital courseware system and the parameters of the student model, the performance and efficiency of the digital courseware system in large-scale applications are guaranteed. By using big data and artificial intelligence technologies, student learning data is continuously collected and analyzed, improving the scientific nature and accuracy of educational decisions. This solves the problems of inaccurate knowledge assessment, difficulty in achieving personalized learning, lack of adaptability, and challenges in large-scale applications that exist in traditional education systems. By adopting a variety of artificial intelligence technologies, efficient and personalized knowledge status assessment and knowledge learning are achieved.

[0136] Suppose that when students are learning the "Algebraic Equations" unit, the system monitors the following behavioral and answer data in real time: Answer performance shows 3 consecutive errors on "Linear Equations in One Variable," with errors concentrated on **handling the transposition symbol** (e.g., incorrectly moving "+3x" to the other side of the equation to become "+3x" instead of "-3x"); 100% accuracy on "Simplifying Fraction Coefficients" problems, but taking a long time (averaging 5 minutes per problem); in problems involving "Multi-Step Integrated Applications," the first attempt fails, but the second attempt succeeds; behavioral data shows multiple viewings of the explanation videos for incorrect questions, with the viewing time exceeding twice the average; frequent switching between knowledge points, especially between "Equation Solving" and "Fraction Operations."

[0137] This application identifies weaknesses by analyzing error sequences using an LSTM model, identifying the "transition symbol rule" as a key weakness, and also finding that students hesitate in "step integration" (which takes too long but has a high accuracy rate).

[0138] This application recommends the following resources: Instant push notifications of symbol rule micro-lessons (2-minute animations + interactive exercises) to strengthen muscle memory for item transfer operations; For the "step integration" problem, step-by-step guided questions are generated (e.g., breaking down multi-step problems into sub-tasks with hints), and "timed challenges" are inserted to improve proficiency; Based on the Q-relevance of the skill mapping matrix, comprehensive cross-knowledge point questions are recommended (e.g., combining fraction operations with equation solving) to solidify knowledge transfer abilities.

[0139] This application also makes path adjustments: a "symbol rules" special training module is temporarily inserted, and the advanced application problems in the original learning path are suspended; if the student's accuracy rate in the next 3 questions increases to 80%, the original path is restored and adaptive difficulty problems are added (such as increasing variable coefficient complexity); if errors still exist, a real teacher intervention mechanism is triggered (such as a pop-up prompting "Do you need one-on-one tutoring?").

[0140] The underlying technical support for this application is: real-time stream processing: through frameworks such as Apache Kafka, data such as answer time, error type, and parsing clicks are collected in real time to trigger model re-evaluation (updating students' knowledge status every second);

[0141] Multi-model collaboration: RNN predicts students' forgetting curves and dynamically adjusts review intervals; collaborative filtering algorithm recommends efficient resources for similar learning groups (e.g., "80% of students who successfully mastered the transfer have used a certain interactive tool"); feedback loop: every time a student completes a recommendation task, the system compares the evaluation results with the actual performance and uses reinforcement learning to optimize subsequent recommendation strategies.

[0142] In a specific implementation of step S104, one embodiment is as follows: the fusion of multiple evaluation results to obtain the evaluation level result includes:

[0143] S1041. Based on the type of the learning task and the relationship between the evaluation results, determine the fusion dimension of the evaluation results;

[0144] S1042. Call the preset fusion method of the fusion dimension to fuse the multiple evaluation results to obtain the evaluation level result.

[0145] In steps S1041-S1042, the learning task set in this application includes single-skill type and multi-skill collaborative type. The resulting evaluation results may exhibit both collaborative and conflicting relationships. Based on the type of the learning task and the relationship between the evaluation results, the fusion dimension of the evaluation results is determined. If the learning task is a single-skill type, there is no need to determine the relationship between the evaluation results; the fusion dimension of the evaluation results is directly determined as the weight allocation dimension. If the learning task is a multi-skill collaborative type, it is necessary to further determine the relationship between the obtained evaluation results. If there is a collaborative relationship between the evaluation results, the fusion dimension of the evaluation results is determined as the collaborative dimension; if there is a conflicting relationship between the evaluation results, the fusion dimension of the evaluation results is determined as the conflicting dimension. That is, this application also sets different fusion methods for different fusion dimensions. When multiple evaluation results are obtained, their fusion dimension is determined, and the preset fusion method for the fusion dimension is called to fuse the multiple evaluation results to obtain the evaluation level result.

[0146] In a specific implementation of step S1042, one embodiment is as follows: the fusion dimension includes a collaborative dimension;

[0147] The step of invoking the preset fusion method of the fusion dimension to fuse the multiple evaluation results to obtain the evaluation level result includes:

[0148] S10421. Based on the knowledge and skills examined in the learning task, determine whether the relationship between the knowledge and skills meets the conditions for collaborative integration.

[0149] S10422. If so, the multiple evaluation results are logically combined to obtain a comprehensive judgment result, and an evaluation level result is obtained based on the comprehensive judgment result.

[0150] In steps S10421-S10422, if the fusion dimension is determined, then the relationship between the knowledge and skills examined in the learning task is judged again to see if it meets the synergistic fusion condition. The synergistic fusion condition is that the knowledge and skills are synergistic. Then, through the integrative skill nodes defined in the skill mapping matrix Q, such as "expression priority processing", the corresponding evaluation results generated by multiple sub-models are logically combined, such as logical AND / OR, to generate a comprehensive judgment result. The comprehensive judgment result is the comprehensive ability judgment result of the student, and the evaluation level result is obtained based on the comprehensive judgment result. For example, if the probability of the student mastering skill A is 0.9 and skill B is 0.8, then the probability of mastering integrative skill C may be the product of the two (0.72). Then the evaluation level result is that the student's knowledge state is 0.72, thus reflecting the synergistic requirement.

[0151] In the weight allocation dimension, the weights of each model's output are dynamically adjusted based on the student model's historical accuracy (e.g., AUC) and real-time data characteristics (e.g., problem complexity). For example, if the LSTM performs better in recent predictions, its weight is increased; if a model consistently overestimates the student's knowledge level, leading to frequent subsequent errors, its weight is automatically reduced. The system continuously collects student feedback (e.g., subsequent answer performance) and compares it with model evaluation results, using reinforcement learning algorithms (e.g., Q-learning) to optimize the model weight allocation strategy.

[0152] In the conflict dimension, if there is a conflict between the evaluation results of multiple sub-models for a certain knowledge skill, such as LSTM predicting a mastery probability of 0.8 and Bayesian network predicting 0.5, then a final decision is made through a preset threshold such as 0.7 or manual intervention rules such as the priority marked by the teacher, so as to obtain the final evaluation level result.

[0153] Example 2

[0154] This application also provides an AI-assisted knowledge evaluation device for adaptive learning, such as... Figure 6 The diagram shows a block diagram of an AI-assisted knowledge evaluation device for adaptive learning. This device performs functions corresponding to the steps described above in executing an AI-assisted knowledge evaluation method for adaptive learning on a terminal device. The device can be understood as a server component including a processor. The AI-assisted knowledge evaluation device for adaptive learning described in this application is applied to a digital courseware system, and the device includes:

[0155] Setting module 601 is used to pre-set learning tasks containing multiple knowledge and skills in the digital courseware system and obtain learning data when students answer the learning tasks.

[0156] The decomposition module 602 is used to decompose the learning task into multiple knowledge skills, map the multiple knowledge skills to the learning data, and filter out key learning data from the learning data to input the key learning data into the student model.

[0157] The evaluation module 603 is used to match the key learning data with multiple sub-models in the student model, so as to perform knowledge status evaluation based on the matched key learning data and obtain the corresponding evaluation results.

[0158] The generation module 604 is used to integrate multiple assessment results to obtain assessment level results, so as to assist in assessing the student's knowledge status based on the assessment level results, and generate a personalized learning path for the student's adaptive learning.

[0159] In one feasible implementation, the decomposition module includes:

[0160] The first module is used to build student response matrices, student profile matrices, and skill mapping matrices based on different types of key learning data and mapped knowledge and skills, respectively.

[0161] The module is used to construct a student model for a student based on the student response matrix, student profile matrix, and skill mapping matrix, and to receive the key learning data after the evaluation is passed.

[0162] In one feasible implementation, the decomposition module further includes:

[0163] The extraction module is used to extract the timestamps of the dynamic data and process the dynamic data based on the timestamps to obtain dynamic sequence data.

[0164] Secondly, based on the dynamic sequence data, student response matrix, student profile matrix, and skill mapping matrix are established respectively.

[0165] In one feasible implementation, the decomposition module also includes:

[0166] The module is used to calculate the correlation between the learning data and the knowledge state using a preset correlation calculation network, and to determine whether the correlation meets preset correlation conditions.

[0167] The input module is used to determine the learning data as key learning data if the condition is met.

[0168] In one feasible implementation, the generation module includes:

[0169] A determination module is used to determine the fusion dimension of the evaluation results based on the type of the learning task and the relationship between the evaluation results;

[0170] The calling module is used to call the preset fusion method of the fusion dimension to fuse the multiple evaluation results to obtain the evaluation level result.

[0171] In one feasible implementation, the generation module further includes:

[0172] The judgment module is used to determine whether the relationship between the knowledge and skills examined in the learning task meets the conditions for collaborative integration.

[0173] The combination module is used to logically combine the multiple evaluation results if the condition is met, to obtain a comprehensive judgment result, and to obtain an evaluation level result based on the comprehensive judgment result.

[0174] In one feasible implementation, the evaluation module includes:

[0175] The partitioning module is used to partition the key learning data into corresponding attribute types according to predefined attribute partitioning relationships;

[0176] The matching module is used to determine the sub-model matched by the key learning data based on the sub-model type of the student model matched by the attribute type.

[0177] Example 3

[0178] This application also provides an electronic device, such as Figure 7 As shown, it includes: a processor 701, a memory 702, and a bus 703. The memory 702 stores machine-readable instructions that can be executed by the processor 701. When the electronic device is running, the processor 701 and the memory 702 communicate through the bus 703. When the machine-readable instructions are executed by the processor 701, the steps of any one of the artificial intelligence-based assisted knowledge assessment methods described above are executed.

[0179] Example 4

[0180] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any one of the artificial intelligence-based assisted knowledge assessment methods described above.

[0181] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0182] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0183] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. If the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0184] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-assisted knowledge evaluation method for adaptive learning, characterized in that, The method, applied to a digital courseware system, includes: Learning tasks containing multiple knowledge and skills are pre-set in the digital courseware system, and learning data is obtained when students answer the learning tasks. The learning task is decomposed into multiple knowledge skills, which are then mapped to the learning data. Key learning data is then selected from the learning data and input into the student model. The key learning data includes dynamic data, which includes the time interval between multiple attempts on the same question and the time trajectory of jumping between different knowledge points. The key learning data is matched with multiple sub-models in the student model to obtain corresponding evaluation results by evaluating the knowledge status based on the matched key learning data. The fusion dimension of the evaluation results is determined based on the type of the learning task and the relationship between the evaluation results. The type includes multi-skill collaborative type; the relationship includes collaborative and conflict relationships; the fusion dimension includes collaborative and conflict dimensions. The dynamic data includes the time interval between multiple attempts on the same problem and the time trajectory of jumping between different knowledge points, which are input into the multiple sub-models, including deep learning models such as RNN and LSTM, for knowledge status evaluation. If the learning task includes multiple knowledge skills, a cognitive diagnostic model based on a skill mapping matrix is ​​also invoked for knowledge status evaluation. Multiple assessment results are integrated to obtain assessment level results, which are used to assist in assessing students' knowledge status and generate personalized learning paths for students' adaptive learning. The input of the key learning data into the student model includes: Based on different types of key learning data and mapped knowledge and skills, student response matrices, student profile matrices, and skill mapping matrices are established respectively. The skill mapping matrix defines the direct association between each learning task and skill, as well as complex tasks that require the collaborative application of multiple skills. The student response matrix is ​​obtained by binarying the number of students and the learning results of the learning tasks. The skill mapping matrix includes integrated skills, which correspond to the assessment level results. Based on the student response matrix, student profile matrix, and skill mapping matrix, a student model is constructed using Boolean matrix multiplication, and the key learning data is received after the evaluation is passed. The fusion dimension includes the collaboration dimension; The step of invoking the preset fusion method of the fusion dimension to fuse the multiple evaluation results to obtain the evaluation level result includes: Based on the knowledge and skills examined in the learning task, determine whether the relationship between the knowledge and skills meets the conditions for synergistic integration; If so, the multiple evaluation results are logically combined to obtain a comprehensive judgment result, and an evaluation level result is obtained based on the comprehensive judgment result; the logical combination is a product operation.

2. The method according to claim 1, characterized in that, The process of establishing student response matrices, student profile matrices, and skill mapping matrices based on different types of key learning data and mapped knowledge and skills includes: Extract the timestamps from the dynamic data, and process the dynamic data based on the timestamps to obtain dynamic sequence data; Based on the dynamic sequence data, student response matrix, student profile matrix, and skill mapping matrix are established respectively.

3. The method according to claim 1, characterized in that, The step of filtering key learning data from the learning data includes: The correlation between the learning data and the knowledge state is calculated using a pre-set correlation calculation network, and it is determined whether the correlation meets the preset correlation conditions. If so, then the learning data is determined to be key learning data.

4. The method according to claim 1, characterized in that, The evaluation level result obtained by integrating multiple evaluation results includes: The preset fusion method of the fusion dimension is invoked to fuse the multiple evaluation results to obtain the evaluation level result.

5. The method according to claim 1, characterized in that, The matching of the key learning data with the sub-models in the student model includes: According to the predefined attribute partitioning relationship, the key learning data is partitioned into the corresponding attribute types; Based on the sub-model type of the student model matched by the attribute type, determine the sub-model matched by the key learning data.

6. An AI-assisted knowledge evaluation device for adaptive learning, characterized in that, The device, applied to a digital courseware system, includes: The setting module is used to pre-set learning tasks containing multiple knowledge and skills in the digital courseware system and obtain learning data when students answer the learning tasks. The decomposition module is used to decompose the learning task into multiple knowledge skills, map the multiple knowledge skills to the learning data, and filter out key learning data from the learning data to input the key learning data into the student model; the key learning data includes dynamic data; the dynamic data includes the time interval between multiple attempts on the same question and the time trajectory of jumping between different knowledge points; An evaluation module is used to match the key learning data with multiple sub-models in the student model to obtain corresponding evaluation results by evaluating the knowledge status based on the matched key learning data; based on the type of the learning task and the relationship between the evaluation results, the fusion dimension of the evaluation results is determined; the type includes multi-skill collaborative type; the relationship includes collaborative and conflict relationships; the fusion dimension includes collaborative dimension and conflict dimension; the dynamic data includes the time interval of multiple attempts on the same problem and the time trajectory of jumping between different knowledge points, which are input into the multiple sub-models, including deep learning models such as RNN and LSTM, for knowledge status evaluation; if the learning task includes multiple knowledge skills, a cognitive diagnostic model based on a skill mapping matrix is ​​also invoked for knowledge status evaluation; The generation module is used to integrate multiple assessment results to obtain assessment level results, so as to assist in assessing the student's knowledge status based on the assessment level results, and generate a personalized learning path for the student's adaptive learning. Decomposition module, including: The first module is used to establish a student response matrix, a student profile matrix, and a skill mapping matrix based on different types of key learning data and mapped knowledge and skills. The skill mapping matrix defines the direct association between each learning task and a skill, as well as complex tasks that require the collaborative application of multiple skills. The student response matrix is ​​obtained by binarying the number of students and the learning results of the learning tasks. The skill mapping matrix includes integrated skills, which correspond to the assessment level results. The module is used to construct a student model for a student based on the student response matrix, student profile matrix and skill mapping matrix using Boolean matrix multiplication, and to receive the key learning data after the evaluation is passed. The generation module also includes: The judgment module is used to determine whether the relationship between the knowledge and skills examined in the learning task meets the conditions for collaborative integration. The combination module is used to logically combine the multiple evaluation results if the condition is met, to obtain a comprehensive judgment result, and to obtain an evaluation level result based on the comprehensive judgment result; the logical combination is a product operation.

7. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of an AI-assisted knowledge evaluation method for adaptive learning as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of an AI-assisted knowledge assessment method for adaptive learning as described in any one of claims 1 to 5.

Images